This dataset contains metadata for 102 hand-curated SAE features from the Gemma-2-2B language model. Each feature is characterised along three primary axes: activation density (how frequently the feature activates, ranging from <0.1% to >0.5% of tokens), vocabulary diversity (semantic breadth from single-word detectors to broad concept clusters), and locality (whether activation is concentrated on individual tokens or distributed across longer text spans). The features are selected to represent all 27 possible combinations of these three axes at low/medium/high levels, with at least 2 representative features per combination. The dataset includes feature types such as literal token detectors, conceptual clusters (e.g., emojis), stylistic registers, structural markers, topics, programming languages, and behavioral patterns (e.g., refusal), designed to test whether context modification methods can generate fluent, high-activating inputs across qualitatively different feature categories.
